# Use AI research to ask better referral questions | SourceX

Canonical: https://refer.sourcex.si/guides/how-to-use-public-ai-research-to-ask-better-referral-questions
Last updated: 2026-11-04

Read a paper for what it measured, not for what it seems to say about demand. A finding about training data can point to a property of company records worth asking about: whether people wrote them, when, and whether they record what actually happened. Turn that property into one question the CEO can answer from memory, write down the answer that would end the conversation, and drop everything the paper never said: buyers, prices, and whether SourceX wants a given company's records. Three such questions per conversation is plenty.

This guide is for VC platform strategists and operating partners who read AI research and talk to CEOs. It starts from a paper's own sentence. If you are starting from a headline claim, the [claim-to-question worksheet](/guides/the-questions-that-turn-ai-hype-into-referral-diligence) fits better; if from a new category's requirements, the [category signal map](/guides/what-a-new-data-category-means-for-your-founder-network) does.

## What a paper can tell a referral conversation, and what it cannot

AI research measures models and datasets. It does not measure companies, buyers or licensing programs. Before translating anything, sort the paper by what it measured.

| Research type | What it measures | Record property it can point to | What it can never tell you |
|---|---|---|---|
| Data-quality and deduplication studies | How repeated or templated text affects models | How much of an archive is people writing in their own words | What any archive is worth |
| Synthetic-data studies | What happens when models learn from generated text | Who wrote the records, and when | Whether any developer wants a given company's records |
| Synthetic-substitute studies | How far generated data can stand in for real data | What generated text cannot supply, such as real outcomes | Whether the substitute works in your sector |
| Evaluation benchmarks | How models are tested on tasks | Whether records show the result of the work | Whether the company holds the rights |
| Architecture and hardware papers | Model design and compute | Nothing about company records | Anything useful to a referral conversation |

A paper in the last row should produce no question at all; a platform team that sends one anyway spends a founder's patience on nothing. Evaluation research has its own treatment in [when evaluation data changes which companies you call](/guides/when-evaluation-data-changes-which-companies-you-call).

## The research-to-question translation guide

Four steps, done in writing, dated.

1. **Quote the finding.** Copy the paper's own sentence, not a summary of it. If you cannot find a sentence that says it, the finding is not in the paper.
2. **Name what was measured, and what was not.** Models, corpora, tasks. Then write the list of things the paper did not study; companies and buyers are always on it.
3. **Derive one record property.** A feature of a company's own records that the finding makes relevant, observable without seeing a file.
4. **Write the CEO question and the ending answer.** One question the CEO can answer from memory, and the answer that would close the topic for this company.

Three translations, built from papers checked on 28 October 2026:

| | Lee et al., deduplication | Shumailov et al., model collapse | Gunasekar et al., phi-1 |
|---|---|---|---|
| The finding | "datasets contain many near-duplicate examples and long repetitive substrings. As a result, over 1% of the unprompted output of language models trained on these datasets is copied verbatim from the training data." | "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear." | Paraphrased: a 1.3B-parameter code model trained on filtered web data plus about 1B tokens of synthetic textbooks and exercises from GPT-3.5 reached 50.6% pass@1 on HumanEval |
| What it measured | Public text corpora and the models trained on them | Models trained on earlier models' output | One code model, on coding benchmarks |
| Not studied | Companies, buyers, prices | Companies, buyers, prices | Companies, buyers, any business domain |
| Record property | How much is written by people versus forms and templates | Who wrote the records, and since when | Whether records hold something a generator cannot invent, such as what actually happened |
| Question for the CEO | "Of the records you'd describe, how much is people writing in their own words, and how much is forms or system messages?" | "Since when have your teams written these records, and has any recent material been drafted with AI tools?" | "When a job or case closes, does the record say how it turned out?" |
| Answer that ends it | "Almost all of it is templates or system-generated." | Not an ending by itself; note it, since SourceX's how-it-works page asks suppliers to confirm data "wasn't created or altered with AI for the sale" | "We log the request, never the outcome." |

The third column is there on purpose. phi-1 is counterevidence: if generated text can teach a model as well as real records in a given domain, the case for licensing that domain's records weakens. Translating it into a question keeps the doubt inside the conversation instead of outside it.

**Method note.** Each row was built from the paper's abstract or a verified excerpt, checked on the date above. The translation measures relevance between a finding and a record type. It does not measure demand, volume, price or whether SourceX runs a program for the records, which are settled only in SourceX's review.

## What never survives the translation

Six things fall out at step 2, every time:

- **Demand.** No paper shows that anyone wants a particular company's records.
- **Price.** Nothing in a training-data paper converts into a payout.
- **A buyer.** Papers name the models they trained, never the companies that will license.
- **Eligibility.** SourceX's screening page checks size at peak, records the company created, rights, an authorizing CEO and relevance to a current program. Funding is not among the checks, and neither is having read the right paper.
- **A program.** SourceX checks relevance to its current programs during review. A research topic is not a program.
- **The paper's own caveats.** If the authors limit their claim to one model size or one domain, the limit travels with the question.

## Three questions, asked through the person who knows the CEO

The research questions come after the published fit facts, never instead of them. The fit facts are: 50 or more full-time employees at peak, contractors left out; years of records the company created itself; rights to license them; a CEO who decides; and location. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.

Then pick at most three research questions, one per row above or your own translations, and follow four rules:

1. **The relationship owner asks.** The person who already talks to the CEO asks in a conversation they would have anyway.
2. **Memory, not files.** Every answer is what the CEO says. No samples, exports or screenshots move, and SourceX's own first step asks the company only a few questions, with no files.
3. **Record the answer as a sentence and a date.** "CEO says most notes are typed by adjusters, since 2012" is a usable record; "looks promising" is not.
4. **An ending answer ends it.** If the CEO gives the answer in your ending row, thank them and close the topic for that company.

Before any introduction, ask whether the company has ever applied to, held an account with or exchanged email with SourceX. Section 3 of the terms treats that company as a duplicate, credited to no later partner.

## A small test before you rely on research-derived questions

Whether these questions improve a conversation is itself unproven. Test it on a dozen companies before you build a platform habit around it, and write the plan first.

| Element | Plan |
|---|---|
| Sampling frame | Twelve companies whose relationship owners agree to ask, listed and frozen before the first conversation |
| Consent | The owner asks the CEO whether a few questions about their records are welcome; a no is recorded, not replaced |
| What is asked | The published fit facts, then the same three research questions, in the same order |
| Variables | For each question: answered, unsure or declined; whether the answer changed the decision to continue or stop; minutes taken |
| Analysis | Count the conversations in which a research question, rather than a fit fact, decided continue or stop; report every company, including those never reached |
| Decision rule | Keep the research questions if they decide at least three of twelve conversations and most are answered; drop any question that mostly draws "don't know" |
| Claims allowed | Descriptive only: which questions CEOs could answer. Not that the questions caused better introductions |

No results from this test have been published. For how SourceX treats results drawn from its own records, see [the partner results research guide](/guides/partner-results-research); for a comparison design that tests a thesis rather than a question set, see [why a network experiment beats a portfolio blast](/guides/why-a-network-experiment-is-better-than-a-portfolio-blast).

## When the research turns against the question

- **Substitutes.** If synthetic data keeps improving in a domain, as phi-1 suggests for code, questions about that domain's records lose their reason. Watch for that before adding questions, not after.
- **Revisions.** Papers get revised and contradicted. A question built on a finding that no longer holds should be retired, and dating each translation makes that easy.
- **Rights decide more than research.** A record type can be relevant on every paper and still belong to the company's clients. Rights are read by the company's counsel and SourceX's rights check.
- **Earlier AI licensing.** Data already licensed for AI training is out: the how-it-works page lists not having licensed it before among the things a supplier confirms.
- **Question fatigue.** No data on how CEOs react to research-derived questions has been published. Three is a ceiling, not a target.

## Next step

> **Refer a company. No cap on rewards.** Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

When a CEO's answers pass the fit facts and none of your ending answers came up, the relationship owner should [register](/register), compare those answers with the published [who qualifies](/who-qualifies) profile and go through the [program terms](/terms) on credit before the company files its application.

Reading research to ask better questions is network work. Judging SourceX as a potential investment is a separate exercise, and anyone doing it can ask for materials through the [SourceX investor page](https://sourcex.si/investors).

This article is for information only and is not an offer to sell or a solicitation of an offer to buy securities. Any offering by SourceX is made only to verified accredited investors through its offering documents.

## Sources

- https://arxiv.org/abs/2107.06499
- https://www.nature.com/articles/s41586-024-07566-y
- https://arxiv.org/abs/2306.11644
- https://sourcex.si/how-it-works
- https://refer.sourcex.si/who-qualifies
- https://refer.sourcex.si/terms
- https://sourcex.si/investors
